New and Improved Classifiers for Fault Diagnosis

Many existing fault classification algorithms use Bayesian approach to fuse different classifiers. One notable problem is that the Bayesian approach cannot distinguish between consistent but low probability data and inconsistent data. Moreover, the Bayesian approach cannot properly handle incomplete and uncertain evidence caused by measurement noise or interferences. Here Intelligent Automation, Inc. (IAI) and its consultant Prof. Henry Leung of the University of Calgary, jointly propose two ways to improve the performance of existing fault classification algorithms. First, we propose to apply the latest development in classification theory known as Support Vector Machines (SVM) to improve the accuracy of classifiers. SVM has several attractive advantages, including global optimal solutions, no over-training problems, explicit bounds on achievable performance, and better recognition performance than most learning systems for a wide range of applications. Second, we propose to apply Dempster-Shafer (DS) theory to fuse several classifier outputs together. DS theory has several advantages such as commutative and associative properties, which means the identification results are independent of the order in which the data is presented. The most important advantage is that DS theory can explicitly take the evidence uncertainty into account. Hence the system will be very robust to noise and uncertainties.